src/mesa/drivers/dri/i965/genX_blorp_exec.c

   1 /*
   2  * Copyright © 2011 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <assert.h>
  25
  26 #include "intel_batchbuffer.h"
  27 #include "intel_mipmap_tree.h"
  28 #include "intel_fbo.h"
  29
  30 #include "brw_context.h"
  31 #include "brw_state.h"
  32
  33 #include "blorp/blorp_genX_exec.h"
  34
  35 #if GEN_GEN <= 5
  36 #include "gen4_blorp_exec.h"
  37 #endif
  38
  39 #include "brw_blorp.h"
  40
  41 static void *
  42 blorp_emit_dwords(struct blorp_batch *batch, unsigned n)
  43 {
  44    assert(batch->blorp->driver_ctx == batch->driver_batch);
  45    struct brw_context *brw = batch->driver_batch;
  46
  47    intel_batchbuffer_begin(brw, n, RENDER_RING);
  48    uint32_t *map = brw->batch.map_next;
  49    brw->batch.map_next += n;
  50    intel_batchbuffer_advance(brw);
  51    return map;
  52 }
  53
  54 static uint64_t
  55 blorp_emit_reloc(struct blorp_batch *batch,
  56                  void *location, struct blorp_address address, uint32_t delta)
  57 {
  58    assert(batch->blorp->driver_ctx == batch->driver_batch);
  59    struct brw_context *brw = batch->driver_batch;
  60
  61    uint32_t offset = (char *)location - (char *)brw->batch.map;
  62    return brw_emit_reloc(&brw->batch, offset,
  63                          address.buffer, address.offset + delta,
  64                          address.read_domains,
  65                          address.write_domain);
  66 }
  67
  68 static void
  69 blorp_surface_reloc(struct blorp_batch *batch, uint32_t ss_offset,
  70                     struct blorp_address address, uint32_t delta)
  71 {
  72    assert(batch->blorp->driver_ctx == batch->driver_batch);
  73    struct brw_context *brw = batch->driver_batch;
  74    struct brw_bo *bo = address.buffer;
  75
  76    brw_emit_reloc(&brw->batch, ss_offset, bo, address.offset + delta,
  77                   address.read_domains, address.write_domain);
  78
  79    uint64_t reloc_val = bo->offset64 + address.offset + delta;
  80    void *reloc_ptr = (void *)brw->batch.map + ss_offset;
  81 #if GEN_GEN >= 8
  82    *(uint64_t *)reloc_ptr = reloc_val;
  83 #else
  84    *(uint32_t *)reloc_ptr = reloc_val;
  85 #endif
  86 }
  87
  88 static void *
  89 blorp_alloc_dynamic_state(struct blorp_batch *batch,
  90                           uint32_t size,
  91                           uint32_t alignment,
  92                           uint32_t *offset)
  93 {
  94    assert(batch->blorp->driver_ctx == batch->driver_batch);
  95    struct brw_context *brw = batch->driver_batch;
  96
  97    return brw_state_batch(brw, size, alignment, offset);
  98 }
  99
 100 static void
 101 blorp_alloc_binding_table(struct blorp_batch *batch, unsigned num_entries,
 102                           unsigned state_size, unsigned state_alignment,
 103                           uint32_t *bt_offset, uint32_t *surface_offsets,
 104                           void **surface_maps)
 105 {
 106    assert(batch->blorp->driver_ctx == batch->driver_batch);
 107    struct brw_context *brw = batch->driver_batch;
 108
 109    uint32_t *bt_map = brw_state_batch(brw,
 110                                       num_entries * sizeof(uint32_t), 32,
 111                                       bt_offset);
 112
 113    for (unsigned i = 0; i < num_entries; i++) {
 114       surface_maps[i] = brw_state_batch(brw,
 115                                         state_size, state_alignment,
 116                                         &(surface_offsets)[i]);
 117       bt_map[i] = surface_offsets[i];
 118    }
 119 }
 120
 121 static void *
 122 blorp_alloc_vertex_buffer(struct blorp_batch *batch, uint32_t size,
 123                           struct blorp_address *addr)
 124 {
 125    assert(batch->blorp->driver_ctx == batch->driver_batch);
 126    struct brw_context *brw = batch->driver_batch;
 127
 128    /* From the Skylake PRM, 3DSTATE_VERTEX_BUFFERS:
 129     *
 130     *    "The VF cache needs to be invalidated before binding and then using
 131     *    Vertex Buffers that overlap with any previously bound Vertex Buffer
 132     *    (at a 64B granularity) since the last invalidation.  A VF cache
 133     *    invalidate is performed by setting the "VF Cache Invalidation Enable"
 134     *    bit in PIPE_CONTROL."
 135     *
 136     * This restriction first appears in the Skylake PRM but the internal docs
 137     * also list it as being an issue on Broadwell.  In order to avoid this
 138     * problem, we align all vertex buffer allocations to 64 bytes.
 139     */
 140    uint32_t offset;
 141    void *data = brw_state_batch(brw, size, 64, &offset);
 142
 143    *addr = (struct blorp_address) {
 144       .buffer = brw->batch.bo,
 145       .read_domains = I915_GEM_DOMAIN_VERTEX,
 146       .write_domain = 0,
 147       .offset = offset,
 148    };
 149
 150    return data;
 151 }
 152
 153 #if GEN_GEN >= 8
 154 static struct blorp_address
 155 blorp_get_workaround_page(struct blorp_batch *batch)
 156 {
 157    assert(batch->blorp->driver_ctx == batch->driver_batch);
 158    struct brw_context *brw = batch->driver_batch;
 159
 160    return (struct blorp_address) {
 161       .buffer = brw->workaround_bo,
 162    };
 163 }
 164 #endif
 165
 166 static void
 167 blorp_flush_range(struct blorp_batch *batch, void *start, size_t size)
 168 {
 169    /* All allocated states come from the batch which we will flush before we
 170     * submit it.  There's nothing for us to do here.
 171     */
 172 }
 173
 174 static void
 175 blorp_emit_urb_config(struct blorp_batch *batch,
 176                       unsigned vs_entry_size, unsigned sf_entry_size)
 177 {
 178    assert(batch->blorp->driver_ctx == batch->driver_batch);
 179    struct brw_context *brw = batch->driver_batch;
 180
 181 #if GEN_GEN >= 7
 182    if (!(brw->ctx.NewDriverState & (BRW_NEW_CONTEXT | BRW_NEW_URB_SIZE)) &&
 183        brw->urb.vsize >= vs_entry_size)
 184       return;
 185
 186    brw->ctx.NewDriverState |= BRW_NEW_URB_SIZE;
 187
 188    gen7_upload_urb(brw, vs_entry_size, false, false);
 189 #elif GEN_GEN == 6
 190    gen6_upload_urb(brw, vs_entry_size, false, 0);
 191 #else
 192    /* We calculate it now and emit later. */
 193    brw_calculate_urb_fence(brw, 0, vs_entry_size, sf_entry_size);
 194 #endif
 195 }
 196
 197 void
 198 genX(blorp_exec)(struct blorp_batch *batch,
 199                  const struct blorp_params *params)
 200 {
 201    assert(batch->blorp->driver_ctx == batch->driver_batch);
 202    struct brw_context *brw = batch->driver_batch;
 203    struct gl_context *ctx = &brw->ctx;
 204    const uint32_t estimated_max_batch_usage = GEN_GEN >= 8 ? 1920 : 1700;
 205    bool check_aperture_failed_once = false;
 206
 207    /* Flush the sampler and render caches.  We definitely need to flush the
 208     * sampler cache so that we get updated contents from the render cache for
 209     * the glBlitFramebuffer() source.  Also, we are sometimes warned in the
 210     * docs to flush the cache between reinterpretations of the same surface
 211     * data with different formats, which blorp does for stencil and depth
 212     * data.
 213     */
 214    if (params->src.enabled)
 215       brw_render_cache_set_check_flush(brw, params->src.addr.buffer);
 216    brw_render_cache_set_check_flush(brw, params->dst.addr.buffer);
 217
 218    brw_select_pipeline(brw, BRW_RENDER_PIPELINE);
 219
 220 retry:
 221    intel_batchbuffer_require_space(brw, estimated_max_batch_usage, RENDER_RING);
 222    intel_batchbuffer_save_state(brw);
 223    struct brw_bo *saved_bo = brw->batch.bo;
 224    uint32_t saved_used = USED_BATCH(brw->batch);
 225    uint32_t saved_state_batch_offset = brw->batch.state_batch_offset;
 226
 227 #if GEN_GEN == 6
 228    /* Emit workaround flushes when we switch from drawing to blorping. */
 229    brw_emit_post_sync_nonzero_flush(brw);
 230 #endif
 231
 232    brw_upload_state_base_address(brw);
 233
 234 #if GEN_GEN >= 8
 235    gen7_l3_state.emit(brw);
 236 #endif
 237
 238 #if GEN_GEN >= 6
 239    brw_emit_depth_stall_flushes(brw);
 240 #endif
 241
 242 #if GEN_GEN == 8
 243    gen8_write_pma_stall_bits(brw, 0);
 244 #endif
 245
 246    blorp_emit(batch, GENX(3DSTATE_DRAWING_RECTANGLE), rect) {
 247       rect.ClippedDrawingRectangleXMax = MAX2(params->x1, params->x0) - 1;
 248       rect.ClippedDrawingRectangleYMax = MAX2(params->y1, params->y0) - 1;
 249    }
 250
 251    blorp_exec(batch, params);
 252
 253    /* Make sure we didn't wrap the batch unintentionally, and make sure we
 254     * reserved enough space that a wrap will never happen.
 255     */
 256    assert(brw->batch.bo == saved_bo);
 257    assert((USED_BATCH(brw->batch) - saved_used) * 4 +
 258           (saved_state_batch_offset - brw->batch.state_batch_offset) <
 259           estimated_max_batch_usage);
 260    /* Shut up compiler warnings on release build */
 261    (void)saved_bo;
 262    (void)saved_used;
 263    (void)saved_state_batch_offset;
 264
 265    /* Check if the blorp op we just did would make our batch likely to fail to
 266     * map all the BOs into the GPU at batch exec time later.  If so, flush the
 267     * batch and try again with nothing else in the batch.
 268     */
 269    if (!brw_batch_has_aperture_space(brw, 0)) {
 270       if (!check_aperture_failed_once) {
 271          check_aperture_failed_once = true;
 272          intel_batchbuffer_reset_to_saved(brw);
 273          intel_batchbuffer_flush(brw);
 274          goto retry;
 275       } else {
 276          int ret = intel_batchbuffer_flush(brw);
 277          WARN_ONCE(ret == -ENOSPC,
 278                    "i965: blorp emit exceeded available aperture space\n");
 279       }
 280    }
 281
 282    if (unlikely(brw->always_flush_batch))
 283       intel_batchbuffer_flush(brw);
 284
 285    /* We've smashed all state compared to what the normal 3D pipeline
 286     * rendering tracks for GL.
 287     */
 288    brw->ctx.NewDriverState |= BRW_NEW_BLORP;
 289    brw->no_depth_or_stencil = false;
 290    brw->ib.index_size = -1;
 291
 292    if (params->dst.enabled)
 293       brw_render_cache_set_add_bo(brw, params->dst.addr.buffer);
 294    if (params->depth.enabled)
 295       brw_render_cache_set_add_bo(brw, params->depth.addr.buffer);
 296    if (params->stencil.enabled)
 297       brw_render_cache_set_add_bo(brw, params->stencil.addr.buffer);
 298 }